Prostate cancer: AI that grades biopsies like the experts

Why does the Gleason grade matter so much?

When prostate cancer is suspected, a doctor takes a biopsy: small samples of the gland. A pathologist, the doctor who studies tissue under the microscope, then checks whether there is cancer and how aggressive it is.

That aggressiveness is expressed by the Gleason grade, grouped into grade groups. It weighs heavily on decisions: simple monitoring, surgery or radiotherapy.

But two pathologists do not always give the same grade to the same biopsy, especially if they are not specialists. AI looked promising, but each system was tested by the team that built it, which can inflate the results.

How did the PANDA challenge work?

To avoid that bias, the researchers ran an open competition called PANDA. In 2020, 1,010 teams made up of 1,290 developers from 65 countries built AI from 10,616 digitized prostate biopsies, meaning slides scanned in very high resolution.

The best systems were then tested blind on biopsies from other hospitals in Europe and the United States that their creators had never seen. The reference was the opinion of panels of uropathologists, specialists in tumors of the urinary and genital organs, with 18 to 34 years' experience on the US panel.

The AI was also compared with general pathologists on part of the biopsies.

What were the results?

Agreement is measured with a score called kappa, where 1 means perfect agreement. On the US and European biopsies, the AI reached 0.862 and 0.868 with the experts, the level of pathologists.

On 70 Dutch biopsies, the AI was closer to the experts (0.876) than 13 general pathologists from 8 countries (0.765). On average, it missed 1.0% of cancers, against 1.8% for those pathologists.

On 237 US biopsies, its agreement was close to that of 20 US pathologists. It missed fewer cancers (1.9% against 7.3%), but it gave more false alarms on biopsies without cancer.

What does this mean for patients?

A reliable AI could act as a second pair of eyes for the pathologist, highlight a suspicious area or help hospitals short of specialists. The aim is a fairer, more consistent grade from one hospital to the next.

It is not routine yet: the authors say these tools now need testing in prospective trials, with real patients followed over time.

What are the limits?

The AI was assessed one biopsy at a time, whereas in practice a pathologist looks at several samples for each patient. Only the most common type of prostate cancer was studied.

Only 15 of the 1,010 teams were selected for validation. And on the US biopsies, the AI more often wrongly flagged tissue without cancer: a human is still essential.